I need some help on parameters of torch.onnx.export(…)
1, dynamic axes.what are backgrounds or application requirements for us to set dynamic axes?
dynamic_axes can be used to specify dimensions with a dynamic shape (i.e. the shape is known at runtime and can change). Usually dynamic shapes are used in a temporal dimension or spatial dimensions.
I’m still not very clear, take temporal as example, if we take time as a dimension of input ,we can write the input tensor as
input_image = [W H T]
where W is width, H is height, T is temporal
dimension of input_image is 3(fixed), the value of T is changed at every frame, in this example ,we should set T as dynamic axis?
What I understand about dynamic shape is changing of dimensions, not value of a specified dimension, e.g.,
at frame 1: input_image = [W H T]
at frame 2: input_image =[W H]
this is called dynamic shape of input, right?
No, usually you refer to a change in the actual shape. I.e. in one iteration the input shape could be [1, 1, 1] in the next it could be [2, 3, 4].
Dropping or adding an entire dimension would usually just fail at a layer input as a specific number of dimensions is expected.
Now I get the point, thanks a lot